Research Square · Nov 17, 2026
Comparative Supervised Machine Learning for Predicting Cardiovascular Disease among Indian Adults 45 Years or Older: Evidence from the LASI Wave 1
By Kamble V, Meitei WB, Dhar M
Background Cardiovascular disease (CVD) remains a leading cause of death in world, where rapid demographic aging and epidemiological transition are intensifying the burden of older adults. Machine learning (ML) offers potential to improve early risk stratification beyond traditional clinical risk scores, yet population-level evidence from nationally representative Indian data remains limited. Our study aimed to develop and validate ML models for predicting CVD among older adults using the Longitudinal Ageing Study in India (LASI) wave 1, comparing seven supervised algorithms and identifying key predictive features. Methods The analytic sample consists of 64,447 older adults. Using stratified sampling, the data were divided into a training set (56%, n = 36,089), a calibration set (14%, n = 9,023), and a held-out test set (30%, n = 19,335). The Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set only, to correct class imbalance. Seven algorithms were evaluated: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), XGBoost, LightGBM, and Multilayer Perceptron (MLP). Test-set performance was assessed using \(\:{AUC}_{ROC}\), \(\:{AUC}_{PRC}\), Brier score, and decision curve analysis, with bootstrap 95% confidence intervals. SHapley Additive exPlanations (SHAP) were used to interpret feature contributions. Results Random Forest showed the best discrimination \(\:{\text{A}\text{U}\text{C}}_{ROC}\)72.8%, 95% CI 72.1–73.6; \(\:{\text{A}\text{U}\text{C}}_{PRC}\) 53.86%, 95% CI 52.5–55.2), with LightGBM and XGBoost following closely (\(\:{\text{A}\text{U}\text{C}}_{ROC}\) 72.6% each). These three ensemble ML algorithms formed a distinct upper tier, about three percentage points above the individual classifiers, Logistic Regression, SVM, Decision Tree, and MLP. Random Forest had the lowest Brier score (0.177), achieving a sensitivity of 61.6% and a specificity of 71.9% at its selected threshold (0.309). Age, having diabetes, higher body mass index, and waist-hip ratio, and having sleep difficulty were identified as the most influential feature variables. Predicted risk rises sharply between ages 58 and 60, then plateaus. Conclusion Our findings suggest that an ensemble ML algorithm, the Random Forest classifier, could help identify adults aged 45 years or older at higher risk of CVD using routinely collected survey data, thereby aiding policy-makers in making evidence-based decisions while implementing accessible chronic disease management services.